[HDFS-15942] Increase Quota initialization threads

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.3.1, 3.4.0
• Fix Version/s: 3.3.1, 3.4.0, 3.2.3
• Component/s: namenode
• Target Version/s: 3.4.0
• Hadoop Flags: Reviewed
• Release Note: The default quota initialization thread count during the NameNode startup process (dfs.namenode.quota.init-threads) is increased from 4 to 12.
• Assignee: Stephen O'Donnell
• Reporter: Stephen O'Donnell
• Created: 01/Apr/21 16:24
• Updated: 27/Jan/24 02:48
• Resolved: 06/Apr/21 13:24

Description

On large namespaces, the quota initialization at started can take a long time with the default 4 threads. Also on NN failover, often the quota needs to be calculated before the failover can completed, delaying the failover.

I performed some benchmarks some time back on a large image (316M inodes 35GB on disk), the quota load takes:

quota - 4  threads 39 seconds
quota - 8  threads 23 seconds
quota - 12 threads 20 seconds
quota - 16 threads 15 seconds

As the quota is calculated when the NN is starting up (and hence doing no other work) or at failover time before the new standby becomes active, I think the quota should use as many threads as possible.

I proposed we change the default to 8 or 12 on at least trunk and branch-3.3 so we have a better default going forward.

Has anyone got any other thoughts?